Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CycleGAN with Better Cycles

Tongzhou Wang, Yihan Lin

TL;DR

This project proposes three simple modifications to cycle consistency, and shows that such an approach achieves better results with fewer artifacts.

Abstract

CycleGAN provides a framework to train image-to-image translation with unpaired datasets using cycle consistency loss [4]. While results are great in many applications, the pixel level cycle consistency can potentially be problematic and causes unrealistic images in certain cases. In this project, we propose three simple modifications to cycle consistency, and show that such an approach achieves better results with fewer artifacts.

CycleGAN with Better Cycles

TL;DR

This project proposes three simple modifications to cycle consistency, and shows that such an approach achieves better results with fewer artifacts.

Abstract

CycleGAN provides a framework to train image-to-image translation with unpaired datasets using cycle consistency loss [4]. While results are great in many applications, the pixel level cycle consistency can potentially be problematic and causes unrealistic images in certain cases. In this project, we propose three simple modifications to cycle consistency, and show that such an approach achieves better results with fewer artifacts.
Paper Structure (19 sections, 6 equations, 5 figures)

This paper contains 19 sections, 6 equations, 5 figures.

Table of Contents

  1. Introduction
  2. CycleGAN
  3. Effects of cycle consistency
  4. Guide training
  5. Regularize
  6. Unrealistic artifacts
  7. Better cycle consistency
  8. Cycle consistency on discriminator CNN feature level
  9. Cycle consistency weight decay
  10. Weight cycle consistency by quality of generated image
  11. Full objective
  12. Experiments
  13. Future work
  14. Parameter tuning
  15. Pretrain and fine-tune discriminators
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Generators quickly learn near-identity mapping at training epoch 3.
  • Figure 2: Generators quickly learn color mapping at training epoch 10.
  • Figure 3: Color inversion effect observed at training epoch 6.
  • Figure 4: Comparison among original CycleGAN, CycleGAN with proposed modifications, and CycleGAN with proposed modifications except weighting cycle consistency by discriminator output on horse2zebra dataset. These images are hand picked from training set.
  • Figure 5: Failure case on horse2zebra dataset.